nanomon 0.3.0

Track LLM usage across DSPy program runs

Nanomon

LLM observability for DSPy applications.

What is Nanomon?

Nanomon is an observability platform for LLM applications built with DSPy. It automatically tracks token usage, costs, and tool calls across your DSPy program runs, giving you full visibility into your AI workflows.

Features

• Automatic LLM token usage tracking
• Cost calculation and analytics
• Tool call tracking with the @nanomon decorator
• DSPy integration (Predict, ReAct, ChainOfThought)
• Run context management for grouping related calls
• Multiple storage backends (API, SQLite)

Quick Start

import dspy
from nanomon import NanomonRunContext, configure_nanomon

# 1. Create context and configure tool tracking
context = NanomonRunContext(default_tags=["production"])
configure_nanomon(context._sink)

# 2. Instrument your LM
lm = dspy.LM(model="openai/gpt-4o-mini")
lm = context.instrument_lm(lm)
dspy.configure(lm=lm)

# 3. Track runs - DSPy callbacks are auto-configured
with context.run(tags=["my-task"], metadata={"version": "1.0"}):
    predictor = dspy.Predict(MySignature)
    result = await predictor.acall(input="Hello")
    # Each DSPy module call is automatically tracked as a separate run

Note: Use .acall() for async DSPy module calls to ensure proper callback integration.

Tool Discovery

Nanomon can automatically discover and sync all @nanomon decorated functions in your codebase:

with context.run(discover_tools=True, tags=["my-task"]):
    # Prints discovered tools and syncs them to nanomon API
    result = await predictor.acall(input="Hello")

Output:

🔍 Discovered @nanomon tools:
  - tools.filesystem.read_file
  - tools.filesystem.write_file
  - tools.web.search_web
✓ Synced 3 tools to nanomon

Tool Tracking

Use the @nanomon decorator to track tool calls within your workflows:

from nanomon import nanomon

@nanomon(name="search_web")
def search_web(query: str) -> str:
    """Search the web for information."""
    # Your implementation
    return results

@nanomon(name="read_file")
async def read_file(path: str) -> str:
    """Read a file from disk."""
    # Your implementation
    return content

Tool calls are automatically recorded with timing and success/error status when executed within a context.run() scope.

Links

• Website: https://nanomon.ai
• Dashboard: https://app.nanomon.ai
• Documentation: https://docs.nanomon.ai

License

MIT